Determination and classification of brain tumors on MR images by deep learning method

Rapid and accurate detection of brain tumors through biomedical images provides early and effective treatment for patients, as well as a good quality of life. In recent years, development of computer hardware and easy access to a large number of relevant data have enabled application of artificial intelligence methods to many areas. In parallel with this, machine learning and deep learning technologies, which are applications of artificial intelligence, have also developed and become widespread. Today, deep learning techniques help specialist doctors to detect anomalies in biomedical images, thus enabling earlier, faster and more accurate diagnosis of diseases or intra-body anomalies. In this thesis, the Mask R-CNN technique, which was created by adding a mask (segmentation) branch to Faster R-CNN architecture with transfer learning models, was applied to accurately segment volumes of three types of brain tumors (meningioma, glioma or pituitary) from neighboring tissues and structures as well as to ensure correct classification of these tumors. Detection, segmentation and classification procedures were carried out with Mask R-CNN under a single roof and without the need for applying image processing techniques on the related data set. In our study, ResNet-50 and ResNet-101 architectures, which were trained with both COCO and ImageNet datasets, were used for the basic network of the Mask R-CNN architecture. The proposed model trainings were started with COCO or Imagenet initial weights. At the end of the test process, it was observed that the classification success (75%) of the ResNet-101 backbone pre-trained on COCO images was higher than the classification success (55%) of the ResNet-50 backbone trained with the same data set. With the ResNet-50 backbone, the classification accuracy on the ImageNet dataset (87%) was higher than the classification accuracy (37%) of the ResNet-101 backbone trained on the same dataset. The IoU averages showing the segmentation success of the models in tumor types were found to be 83% for meningioma tumors, 58% for glioma tumors, and 75% for pituitary tumors. The portability of the Mask R-CNN architecture to systems in the healthcare industry will help specialists to detect brain tumors early.